Skip to content
Rafael Patron

Note Notes 2026

AI Act compliance checklist 2026: 30 operational points for companies

Translated from the Italian original.

In July 2026 the EU AI Office published a summary document on the most frequent findings from the first inspections coordinated with national authorities. Of 47 procedures opened in the first six months of full application, 68% showed shortcomings in mandatory technical documentation, 54% in the management of the incident register, 41% in the inventory of AI systems in use. These figures say one precise thing: AI Act compliance, more than extraordinary skills, requires method and operational consistency.

This article is a concrete checklist. Thirty points, grouped by area of intervention, each with a reference to the applicable article of the AI Act.

How to use this checklist

The thirty points are designed for a mid-sized Italian company that uses AI systems developed in-house or supplied by third parties, in administrative, commercial or operational areas. Companies that directly develop high-risk AI systems have additional obligations not fully covered here, in particular the conformity assessment of Article 43 and the CE marking of Article 48. Those scenarios require dedicated specialist advice.

The checklist should be completed at least once a year, ideally with a quarterly review of the active monitoring points. Ownership of the process must be clearly assigned, typically to the AI Officer or to a DPO whose remit is extended to AI. Evidence of completion must be archived so it can be retrieved in case of inspection by the competent authorities, which for Italy means primarily AgID for technical matters and the Garante Privacy (the Italian Data Protection Authority) for personal data processing.

For those building the framework from scratch, I suggest pairing the checklist with the guide to the AI Act applied to digital marketing, which covers in operational detail the systems most widespread in commercial functions.

Governance and roles (points 1-6)

Governance comes first because without clear roles none of the subsequent activities can be completed sustainably.

1. Formally appoint a company AI Officer, possibly combined with another role, with a written mandate and direct access to the board or equivalent management. Reference: general principle of accountability, Article 26 for deployers of high-risk systems.

2. Update the mandate of the existing DPO to include coordination with the AI Officer on matters that intersect with personal data processing. Reference: Article 26, point (g) (cooperation with competent authorities).

3. Establish an AI Steering Committee meeting at least quarterly, composed of the AI Officer, DPO, CISO, head of legal, business representative. Formal minutes retained.

4. Define a written company AI policy, approved by management, clarifying permitted use, prohibited use, the authorization process for new systems, rules on individual shadow AI.

5. Map RACI responsibilities (Responsible, Accountable, Consulted, Informed) for each AI system used in the organization, updated every six months.

6. Define an annual operating budget dedicated to AI compliance, separate from the general IT budget. Without a dedicated budget, compliance does not get completed, even with the best declared intentions.

Inventory and classification (points 7-12)

The inventory is the factual basis for everything else. Authority inspections always start here.

7. Compile and keep up to date an AI System Inventory listing all AI systems in use: name, supplier, version, purpose, data processed, business owner, date of first activation. Spreadsheet or dedicated tool, what matters is that it is retrievable.

8. Classify each system according to the AI Act categories: unacceptable risk (prohibited), high risk (Annex III), limited risk (transparency obligations), minimal risk. Reference: Articles 5, 6, 50, 52.

9. For each system, check whether it relies on a GPAI model as its underlying model (e.g. use of GPT, Claude, Gemini via API) and document the GPAI provider with its date of conformity to Article 53.

10. Identify the systems that involve personal data processing and link each to the GDPR record of processing activities under Article 30 of Regulation (EU) 2016/679.

11. For systems classified as high-risk, verify the presence of the provider’s technical documentation required by Annex IV and archive a copy or a reference to access. Reference: Article 11.

12. For systems purchased or used as a service, retain contractual evidence of the supplier’s responsibility for AI Act conformity, ideally in a dedicated clause of the service contract.

Transparency toward users and employees (points 13-17)

Transparency is a cross-cutting obligation and produces evidence that authorities can easily verify.

13. Inform customers and prospects when they interact with conversational AI systems, clearly and accessibly before the interaction begins. Reference: Article 50(1).

14. Label content generated or significantly modified by AI when published on company channels, including images, video, audio. Adopt technical standards such as C2PA or SynthID where possible. Reference: Article 50(2)-(3).

15. Inform employees about the use of AI systems in work activities that concern them, with particular reference to systems affecting evaluation, monitoring, work organization. Reference: Article 26(7) and Article 4 of the Italian Workers’ Statute in the version coordinated with transparency rules.

16. Update the privacy notice under Articles 13 and 14 GDPR to include processing involving AI, mentioning the logic involved, the significance and the envisaged consequences under Article 22.

17. Publish on the company website a concise AI transparency page listing the categories of systems used in dealings with the public. It is not expressly mandatory, but it is a practice authorities appreciate and it improves the trust relationship with customers.

DPIA and impact assessment (points 18-21)

Impact assessment is mandatory for high-risk systems and recommended for other sensitive scenarios.

18. Conduct a DPIA under Article 35 GDPR for every AI system that processes personal data on a large scale or significantly affects individuals. Document methodology, outcomes, mitigation measures adopted.

19. For high-risk AI systems, supplement the DPIA with the Fundamental Rights Impact Assessment provided for by Article 27 of the AI Act, separately documented. Reference: Article 27(1)-(5).

20. Update DPIAs at least annually or at every substantial modification of the system. Outdated DPIAs are not acceptable evidence in an inspection of modified systems.

21. Keep the DPIA register as an accessible repository, classified by system type, date of last review, responsible owner. Reference: accountability principle, Article 24 GDPR.

Training data, quality, bias (points 22-25)

For AI systems developed in-house or customized through fine-tuning, training data quality is subject to a specific AI Act obligation.

22. Document the provenance, composition and quality of the datasets used to train or calibrate AI systems developed in-house. Reference: Article 10 for high-risk systems, general accuracy principle of GDPR Article 5, point (d) for personal data.

23. Verify the representativeness of datasets with respect to the system’s target populations, with statistical analysis of distributions by sensitive variables (gender, age, geographic area, where relevant). Document the analysis and any mitigations.

24. Conduct periodic testing to detect discriminatory bias in the system’s results, on groups protected by Article 21 of the EU Charter of Fundamental Rights. Document methodology, frequency, acceptability thresholds.

25. Keep logs of the input data received by the system in production and of the results produced, at least for high-risk systems, for a minimum period allowing retrospective verification and possible review. Reference: Article 12 and Article 19.

Monitoring and incidents (points 26-30)

Continuous monitoring is an obligation under Article 72 and requires tools, procedures, archived evidence.

26. Implement operational monitoring of AI systems in production with metrics defined in advance (accuracy, latency, data drift, quality indicators specific to the use case). Minimum frequency weekly for high-risk, monthly for other systems.

27. Define a formalized AI incident management procedure with response times, escalation channels, recording forms. Reference: Article 73, obligation to notify serious incidents to the competent authorities within 15 days of discovery.

28. Keep a separate, searchable AI incident register, with severity classification, system involved, mitigations applied, outcome. Minimum retention five years.

29. Define and test at least annually the emergency shutdown procedure for critical AI systems, with documented owners and timings. Inspections always check for operational kill switches.

30. Schedule an internal or external audit of AI Act compliance at least annually, with a formal report to management and a remediation plan for the non-conformities found. External audit is recommended from the third year of application of the Regulation onward.

How to size the effort

For an Italian company with 50-200 employees using about ten AI systems spread across HR, marketing, customer service and back office, initial completion of the checklist typically requires four to six months of distributed work, with a commitment of one FTE equivalent for the setup phase and 0.3-0.5 FTE for subsequent maintenance. Additional direct costs (specialist legal advice, a possible AI governance platform, AI literacy training) fall in an indicative range of 20,000-50,000 EUR in the first year, 8,000-20,000 EUR at steady state.

Figures that may look high for an SME, but they must be compared with the potential costs of not acting. AI Act penalties reach 35 million EUR or 7% of global turnover for the most serious violations. Even excluding the worst-case scenario, a non-compliance challenge slows company digitalization, calls into question contracts with customers that require supplier compliance guarantees, and can preclude access to public tenders where AI compliance is starting to be a requirement. Companies that invested early in AI governance are turning the regulatory obligation into a competitive advantage in tenders and in relationships with enterprise customers.

Training resources and networks

Internal competence is the starting point, but it is rare for an Italian company under 500 employees to have all the necessary skills in house. Integration with specialist consultants and professional networks is an operational component of the model, not a shortcut.

For staff AI literacy training, the European reference consists of programs that issue recognized European digital credentials, according to the EU Commission’s EDC framework. In Italy, AIPIA is the first association authorized to issue European credentials for AI training to its members and to those who complete affiliated training programs. For those who want to complete the compliance picture with staff training, the reference address is aipia.it/credenziali-europee/, where active programs and admission requirements are published.

For monitoring regulatory developments, the reference is the enterprise summary of the Stanford AI Index 2026, which tracks with quantitative data AI adoption in companies and the evolution of the international regulatory framework, useful for comparative benchmarking with sector peers.

Frequent mistakes to avoid

Five error patterns observed regularly in companies that start the compliance journey without a structured method. Underestimating the inventory: without an inventory nothing gets applied, yet it is the point companies treat most superficially at the start. Dedicating two full weeks to the inventory in the first quarter is an investment recovered tenfold in the following months.

Confusing the AI Officer with the DPO: they are compatible roles with distinct skills; combining them is viable only if the person actually has both skill sets and enough time. Underestimating the time required: AI Act compliance is not a fixed-term project, it is an ongoing function; companies that treat it as a closed project find themselves out of compliance within 12-18 months.

Outsourcing everything: consultants help in the setup phase, but knowledge of one’s own systems must be internal; authorities ask for explanations from those who operate, not from those who advise. Thinking training is enough: without processes, controls and archived evidence, staff training does not constitute demonstrable compliance; it is a necessary but not sufficient prerequisite.

Where to start if you are behind

Many Italian companies find themselves in 2026 having completed fewer than half of these thirty points. The situation is recoverable, although it requires clear priorities in the first three months of intervention. The first month goes to appointing the AI Officer, the system inventory, AI Act classification, identification of missing DPIAs. The second month to supplier contract documentation, notices to customers and employees, a first draft of the company AI policy. The third month to the incident procedure, operational monitoring of the most critical systems, the AI literacy training plan.

Once these three months are complete, the company has covered the points authorities focus on in a first inspection and has enough solidity to calmly tackle the progressive completion of the remaining elements. Perfect compliance does not exist; reasoned, documented, constantly evolving compliance does, and it is the kind authorities recognize as acceptable in the supervision phase.

In the same section

The whole section

Advisory, talks and contact

Contact Rafael

Advisory, speaking and contact requests are handled on rafaelpatron.it.